Apple is evaluating technology from startup PrismML to compress artificial intelligence models so they can run directly on iPhone hardware [1].
This move represents a critical shift toward on-device processing. By reducing the reliance on cloud-based servers, Apple aims to increase the speed of AI responses, lower operational costs, and enhance user privacy for features like Siri [1, 2, 3].
PrismML, which is backed by Khosla Ventures, specializes in model compression [1, 2]. According to reports, the startup's technology can reduce the memory usage of AI models by up to 15 times [1]. This efficiency would allow the iPhone to host significantly larger models than previously possible without compromising the device's performance or battery life [3, 4].
Local execution of AI is a primary goal for the company as it competes with other mobile operating systems to integrate generative AI. Processing data on the device ensures that sensitive user information does not leave the hardware, a core pillar of Apple's marketing strategy [2, 3].
While Apple has not officially announced a partnership, the company is assessing how PrismML's breakthroughs could be integrated into future software updates or hardware iterations [1, 2]. The ability to run the largest local AI models ever seen on an iPhone would potentially eliminate the latency associated with sending requests to a remote data center [3, 4].
Industry analysts said that the current bottleneck for mobile AI is the limited RAM available on smartphones. The reported memory reduction [1] could effectively bypass these hardware constraints, allowing more sophisticated reasoning and language capabilities to function offline [3].
“Apple is evaluating technology from startup PrismML to compress artificial intelligence models.”
The potential integration of PrismML technology suggests Apple is prioritizing 'edge AI' to gain a competitive advantage in privacy and latency. If Apple successfully deploys models that are 15 times more memory-efficient, it could decouple advanced AI capabilities from expensive cloud infrastructure, making high-performance AI a standard feature of the hardware rather than a subscription-based cloud service.
